Detection Of Abnormal Crowd Events In High Density Video

With the development of computer vision, intelligent video surveillance has been developed more widely. Especially with the improvement of people’s consciousness of public security, people have paid more and more attention to the abnormal behavior detection research in crowed scenes. In this thesis, we focus on the abnormal behavior detection of crowed scene in intelligent video surveillance system, analyze several common crowed abnormal behaviors, and put forward some relative anomaly detection method. As we study the abnormal behavior of high density scene, we should classify the scene crowed density firstly, then detect the abnormal events in the high density crowed scene.For the crowd density estimation problem, this thesis uses the method of least squares linear fit and the method based on image texture analysis. Linear fitting method, use foreground detection method to extract foreground pixels number of targets, count the number of population in the crowed scene artificially, use the method of least squares linear fit, this method has low computational complexity, but the method can’t be used in the crowd scenes where collision happens. Crowd density estimation method based on image texture information reflects the different density crowd, Firstly it extracts the crowd foreground object, and then calculates the foreground object of the gray level co-occurrence matrix, using support vector machine to classify density. This method can be used in the crowd scenes where collision happens. As this method just calculates foreground objects texture information, excludes the influence of the background, the correct rate of classification results is reliable.An efficient anomaly detection technique is proposed based on the cell speed as well as the number of foreground pixels and the movement direction. It is capable of dealing with crowded scenes where the traditional tracking based approaches tend to fail. Initial foreground segmentation of the input frames confines the analysis to foreground objects and effectively ignores irrelevant background dynamics. Input frames are split into non-overlapping cells, followed by extracting motion features by computing the optical flow of the foreground pixels to discriminate the normal and the Anomaly. Speed characteristics can detect high-speed abnormalities. In order to distinguish between vehicle and large object formed by people walking close to each other, this thesis puts forward the statistical method of the direction of movement. Experiments demonstrate that the proposed method has better detecting effect in a short period of time. The fighting of a large number of people and people gathering event can be detected by characteristics such as crowed density, motion intensity, direction of motion and crowed area change. We also can predict the things after people gathering event happens.